When people discuss AI models like GPT-4, Llama, or Mistral, you’ll often hear them described as “open-source” or “closed-source” — a distinction that has nothing to do with how capable a model is, but a lot to do with who can see it, modify it, and run it. Here’s what that distinction actually means in practice.
What Makes an AI Model Closed-Source?
A closed-source AI model is one whose underlying weights, training code, and architecture details are kept private by the company that built it. Users interact with the model only through an interface or API the company controls — a chat app, a developer API — without ever gaining direct access to the model file itself. Most of today’s most well-known consumer AI products, including the models behind many leading chatbots, are closed-source: the company runs the model on its own infrastructure, and the public never downloads or directly inspects the model.
What Makes an AI Model Open-Source?
An open-source (or more precisely, “open-weight”) AI model is one whose trained weights — the actual numerical parameters described in our article on model weights and training — are released publicly, allowing anyone to download the model, run it on their own hardware or cloud infrastructure, inspect how it behaves, and in many cases modify or fine-tune it for their own purposes. Several major AI labs and research organizations have released open-weight versions of their models alongside or instead of closed, proprietary ones.
It’s worth noting that “open-source AI” is used somewhat loosely in the industry. True open-source, in the traditional software sense, would mean the training code, training data, and methodology are all public too — not just the final trained weights. Many “open” AI models release only the weights, which is why you’ll sometimes see the more precise term “open-weight” used instead.
Key Differences in Practice
- Where it runs: closed-source models run exclusively on the provider’s servers; open-weight models can be run on your own hardware, a cloud server you control, or even, for smaller models, a capable personal computer.
- Customization: open-weight models can be directly fine-tuned and modified by anyone with the technical skill and computing resources to do so; closed-source models typically only offer whatever customization options the provider chooses to expose, like limited fine-tuning through an official API.
- Transparency: open-weight models let researchers and developers directly inspect and test the model’s behavior; closed-source models are effectively a “black box” from the outside, with behavior known only through observed outputs.
- Cost structure: closed-source models are typically accessed through usage-based pricing (pay per use via an API or subscription); open-weight models have no per-use licensing fee, though running them still requires computing resources, which have their own real costs.
- Data privacy: running an open-weight model on your own infrastructure means your data never leaves your own systems, which matters for some privacy-sensitive or regulated use cases. Closed-source models process your input on the provider’s servers, governed by that provider’s data policies.
Why Companies Choose One Approach or the Other
Companies keep models closed-source for several reasons: protecting a competitive advantage built through expensive training runs, maintaining tighter control over how the model is used and presented, and reducing the risk of a powerful model being used irresponsibly once it’s out of the company’s hands. Companies release open-weight models to build developer goodwill and ecosystem adoption, accelerate research and external scrutiny of the model’s behavior, and, in some cases, as a strategic move against competitors who rely on closed, paid access as their primary business model.
Which Approach Is “Better”?
Neither approach is universally better — it depends entirely on the use case. A business that wants the easiest possible setup, doesn’t want to manage its own infrastructure, and is comfortable with usage-based pricing will often prefer a closed-source model accessed via API. A business or developer that needs full control over data privacy, wants to deeply customize model behavior, or wants to avoid ongoing per-use fees at scale may prefer an open-weight model they can host themselves. Many organizations end up using a mix of both, depending on the specific task.
Why This Matters for Evaluating AI Software
When evaluating any AI-powered tool, knowing whether it’s built on an open-weight or closed-source model can be genuinely useful context: it affects where your data is processed, how much the vendor can customize the underlying model for their specific product, and what happens to your access if the vendor’s business or pricing changes down the line.
The Bottom Line
Closed-source AI models keep their weights and internals private, accessible only through a controlled interface or API, while open-source (or open-weight) models release their trained weights publicly for anyone to download, run, and modify. The distinction shapes cost, customization, transparency, and data privacy in meaningfully different ways — and understanding it helps clarify what you’re actually getting when you choose one AI tool over another.